Aegis Gateway 2.0

Validation report

This report records the builds, target versions, and adopter runs for Aegis Gateway 2.0. Individual claims are in evidence/CLAIMS-MATRIX.md. Each adopter page names the target version that was exercised.

Tested environments

The local runs in this report and in adopter status used one Windows workstation:

Part Version Used for
Operating system Windows 11, x64 workstation every local run
JDK 17.0.17 or 21.0.11 (Microsoft build) the Maven build, the tests and the gateway jar
JDK 21.0.11 to 21.0.12 Keycloak only (token-exchange evidence)
Maven 3.9.11 the build
Python 3.13 Python clients, Python engines and tools
Node.js 24.14.0 the stdio MCP filesystem server
Apache Knox 3.0.0 the Knox front door
Apache ZooKeeper 3.9.3 ZooKeeper converter and Knox front-door run
Apache Kafka 3.9.1 kafka-admin source and Kafka front-door run
Apache Iceberg 1.10.1 REST catalog front-door run
Apache Flink 1.20.1 RealFlinkTest
Apache Spark 4.0.1 RealSparkTest, RealSparkConnectServerTest
Apache Derby 10.16.1.1 JDBC converter and derby-local state
PostgreSQL 17.6 two gateway processes (evidence)
OpenBao 2.7.1 KV secret provider (evidence); HashiCorp Vault 1.17.6 dev also exercised 2026-10-05 (same KV v2 API)
Keycloak 26.8.0 RFC 8693 token exchange (evidence)
a2a-sdk (Python) 1.2.1 the A2A agent behind the a2a source (A2aAgentFrontDoorsTest, both front doors) (recorded, no saved results)
Helm, kubeconform, hadolint, actionlint 4.3.0, 0.8.0, 2.15.1, 1.7.12 chart, manifest, Dockerfile and workflow checks (chart README)
Apache CouchDB 3.4.2 (official Docker image) CouchdbFrontDoorsTest own host and Knox (adopter status)
Apache Pinot 1.2.0 (official Docker QuickStart) PinotFrontDoorsTest own host and Knox
Apache Hive 4.0.1 (official HiveServer2 image) HiveFrontDoorsTest own host and Knox, 2026-10-11
Apache Ranger 2.9.0 (official Admin and database images) RangerFrontDoorsTest own host and Knox, 2026-10-11

The temporary external PostgreSQL run is described in the PostgreSQL validation evidence; its engine version was not retained. The engine version of each adopter run is in examples/adopters/README.md. Docker was used locally on 2026-10-05 for PostgreSQL 17.6, Vault 1.17.6 dev, and the gateway image smoke test (E2E validation record). Helm 4.3.0 and kubeconform 0.8.0 were installed locally for chart checks; no Kubernetes cluster install was performed. No Windows Subsystem for Linux distribution was installed. Docker on this Windows host ran the catalog’s container adopters, including Airflow, CouchDB, Doris, Druid, Hadoop, HBase, Hive, NiFi, Pinot, Ranger, Storm, and Superset. Atlas, Cassandra, Impala, and Paimon still have no product process here.

Default build (2026-10-03)

$ mvn -B "-Prelease" clean verify
BUILD SUCCESS          20 reactor projects, 364 tests, 0 failures, 0 errors, 0 skipped
RAT                     0 unapproved files
Release profile         source jars and CycloneDX SBOMs generated under target/

The default reactor covers governance, configuration, and packaging. Opt-in real-engine runs are recorded in the sections below. PostgreSQL shared-state runs are in the dated integration evidence.

Full matrix revalidation (2026-10-05)

See evidence/E2E-VALIDATION-2026-10-05.md for the historically reported claim-by-claim pass on that mutable working-tree snapshot, including ZooKeeperFrontDoorsTest (Knox bundle SHA-256 0726C4CE71D23856F8FE9A58BE2C80E7FC7E4AE740ED70E8ECC8CBFB2617B780), Keycloak token exchange, Helm/kubeconform/start_rendered.py, hooks e2e, and Docker Postgres/Vault ITs. The later temporary external PostgreSQL run was not part of this date’s record.

Security control validation (2026-10-05)

See evidence/SECURITY-VALIDATION.md for the focused S1 to S31 control-to-test plan and its 227-test result. See the product threat model for the control boundaries.

Runs added on 2026-10-03

Run Result Against
A2aAgentFrontDoorsTest (-Preal-engines, AEGIS_A2A_PYTHON) 2 of 2 passed, 0 skipped (recorded, no saved results) an agent on the official a2a-sdk 1.2.1, own host and stock Knox 3.0.0
OpenApiDraftIcebergTest (-Preal-engines) 1 of 1 passed the Iceberg 1.10.1 REST document and the official RESTCatalogServer
examples/adopters/token-exchange/keycloak_e2e.py 5 of 5 checks passed Keycloak 26.8.0 (evidence)
adopters/hooks/e2e_gateway.py 7 of 7 checks passed the gateway jar as a process, CPython http.server, the three reference hooks
adopters/hooks/test_hooks.py 12 of 12 passed the hook services over real HTTP
Helm chart checks lint 0 failures, kubeconform 7 of 7 valid, rendered configuration reached /readyz chart README

Signed server catalog and break-glass against real engines (2026-10-06, local working tree)

mvn -B -Preal-engines -pl examples/aegis-validation verify -Dtest=FlinkFrontDoorsTest,IcebergFrontDoorsTest,SparkFrontDoorsTest,KafkaFrontDoorsTest with AEGIS_KNOX_HOME set: 17 of 17 passed, 0 skipped. That count includes the existing two-door scenarios, which rerun unchanged on the modified harness.

For each engine, SignedServers reads the definitions with REQUIRE_PINS (nothing trusted), signs a server catalog of every configured source with its configuration digest and exact pins, and opens it behind the own host and a stock Knox 3.0.0 gateway. GET catalog must list exactly the signed tools as ACTIVE, with the signed digests, and a real read must succeed. A break-glass order is then signed with a separate incident key.

Engine Signed catalog Break-glass order Result through both doors
Flink 1.20.1 MiniCluster flink flink=QUARANTINED catalog read of a real job; with the only server contained, start-up is refused (Knox log: no source that the signed server catalog lets start); an edited baseUrl is refused on the own host
Iceberg 1.10.1 RESTCatalogServer iceberg iceberg=REVOKED real read and resource-scope refusal under the catalog; contained start-up refused
Spark 4.0.1 application UI + standalone Master spark-history, spark-master spark-master=REVOKED Master tools NOT_EXPOSED to reader and operator, the waiting driver stays SUBMITTED, application reads still governed
Kafka 3.9.1 KRaft broker + Connect worker kafka-connect, schema-registry, kafka-admin kafka-connect=REVOKED Connect tools NOT_EXPOSED, the real connector is not deleted, broker reads still governed

Every audit journal from the run was afterwards checked by hand with --verify-audit (2026-10-07: 12 of 12 journals OK); the harness itself does not run that check. The run records stay under examples/aegis-validation/target/adopter-results/<engine>-server-catalog and <engine>-break-glass: transcripts, the signed catalog, the order, public keys and Knox logs. They are not copied into examples/adopters/. The catalogs were built with the library API (SignedServerCatalog, SignedCatalogManifest), not the CLI; the CLI path is covered on the own host by ServerCatalogFlowTest. Spark Connect and Spark with Iceberg (module aegis-validation-spark-connect) were not run with a catalog.

Dated opt-in real-engine profile (2026-10-05)

The real-engine examples are intentionally not part of the default build. On 2026-10-05 the catalog marked 26 adopter/scenario profiles as real two-door passed in examples/adopters/README.md (the earlier 2026-10-03 cohort plus the Docker-backed Airflow, CouchDB, Doris, Druid, Hadoop, HBase, NiFi, Pinot, Storm, and Superset Keeps, and the spark-iceberg scenario). Each passed Keep was run through both own-host Aegis and the then-built knox-bundle jar in a stock Knox 3.0.0 process. Retained folders were refreshed from those runs; the per-project tutorials in TUTORIALS.md are built from them. That 26 is the catalog on 2026-10-05. The current catalog is 28 retained runs: 27 adopter profiles plus the Spark and Iceberg tutorial scenario. Atlas, Cassandra, Impala, and Paimon are the 4 configuration-only profiles. Hive and Ranger were added on 2026-10-11. Each retained folder contains:

The retained materials record the own-host and Knox paths that were exercised. The claims matrix is authoritative for whether each run remains current after later code or configuration changes; a release candidate still needs its own run.

The other 6 shipped profiles (Atlas, Cassandra, Impala, Paimon, AsterixDB, SkyWalking) are config/contract supported as configuration; the service prerequisite for each is recorded in the same adopter index and on ROADMAP.md.

Inventory contract

AdapterConfigsTest, Coverage01VsTwoTest and DocumentationClaimsTest make the public inventory executable:

Claim Current value How it is checked
Shared reference configurations 7 executable reference profiles plus a separate teaching template one executable YAML and developer README in each adopters/configurations/<profile>/ directory; DocumentationClaimsTest rejects flat root YAML and missing catalog links
Adapter configurations 27 deployable adapter configurations one executable YAML in each adopters/configurations/adapters/<adopter>/ directory
Adapter tools 123 tools summed from the same parsed YAML
0.1 frozen inventory 156 tools, 33 engines checked against the shipped 2.0 configuration
Implemented 0.1 successor surface 152 tools exact frozen verdict accounting
New Kafka Admin capabilities 2 tools real AdminClient cluster and configuration reads
Full comparison surface, including Apache Burr (incubating) 165 tools exact Coverage01VsTwoTest assertion
Planned later 5 engines Arrow, Beam, Hudi, Kudu, and Ozone are typed sources for a later release
Less-guarded implemented 0.1 tools 0 class-order assertion

The 27 deployable adapter configurations and 123 tools are governed in the current default build. Their external compatibility state is below and in the claims matrix.

Test boundaries: real software versus deterministic fixtures

No production configuration selects a test-double backend. Non-production fixtures stay in src/test and support narrow assertions only:

Fixture Scope What it proves What it cannot prove
TestMcpServer, TestStdioMcpServer test sources, published only in the Maven test jar protocol negotiation, sessions, pagination, SSE, bounds, content governance and failure handling independent of the production handler interoperability with every external MCP implementation
RestRouteContractHarness examples/aegis-examples/src/test configured method/path/body rendering and governance order for the 27 shipped configurations that a target product/version serves the route; a real-target run remains required
local HTTP endpoints in OPA, JWKS, hook and OTLP tests module test sources exact request/response, timeout and fail-closed behavior compatibility with a named external product unless a separate real-target row says so

Two in-memory classes are production code but are not backend simulations. InMemoryStateStore is the real, bounded, one-process default governance state with no restart durability. InMemoryAuditSink is a programmatic embedding/test observer; configured gateways always create HashChainFileAuditSink as the primary journal. Embedded Apache Derby and ZooKeeper tests run actual upstream implementations, not hand-written protocol fixtures.

Evidence by target

The detailed target-status source is examples/adopters/README.md: it distinguishes 28 retained runs (27 adopter profiles plus the Spark and Iceberg tutorial scenario) from 4 configuration-only profiles (Atlas, Cassandra, Impala, Paimon). The table below is supplemental source/converter detail.

Target Recorded evidence Qualification
JDBC converter 16 tests against Apache Derby 10.16.1.1, including result cap, error mapping and read-only behavior recorded local converter behavior
ZooKeeper converter 16 tests against ZooKeeper 3.9.3 through Curator TestingServer, including digest-authenticated owner-only node creation and an unreachable ensemble counted as an availability failure; AdminServer command names resolve through ZooKeeper’s own registry recorded local converter behavior
Flink 1.20.1 RealFlinkTest starts a real Flink MiniCluster and verifies read, denied cancel, dry-run, approval and real cancellation against the canonical config recorded local source behavior; rerun when Flink source, Flink version or config changes
Spark 4.0.1 RealSparkTest starts a real local Spark driver and verifies History/driver-UI reads through the canonical config; standalone-master submission/kill remains unexercised recorded local read-source behavior; standalone-master write path still needs its own run
Spark Connect 4.0.1 RealSparkConnectServerTest starts a real Spark Connect server in a separate JVM and exercises reads plus a dry-run, independent approval and real destructive drop recorded local source behavior; rerun when Spark Connect source, Spark version or config changes
Iceberg 1.10.1 REST catalog IcebergFrontDoorsTest: real RESTCatalogServer, own host and Knox, recorded 2026-10-05 recorded local source behavior; the release-candidate repeat is on the roadmap
Kafka 3.9.1 and Connect KafkaFrontDoorsTest: real KRaft broker and Connect worker, own host and Knox, recorded 2026-10-05; every default build also drives the kafka-admin source against a real broker recorded local source behavior; the release-candidate repeat is on the roadmap
Atlas, Cassandra, Impala, Paimon configuration and AdapterConfigsTest governance flow product-stack runs are on the future roadmap
Arrow, Beam, Hudi, Kudu and Ozone planned typed sources roadmap
aegis-state-postgres PostgresStateStoreIT: five tests for shared nonce and rate state, persistence, expiration, escaped scans, invalid rate requests, environment-named credentials, and signed-serial high-water rollback PostgreSQL validation evidence; the release-candidate repeat is on the roadmap
Two gateway processes, one postgres namespace TwoGatewayProcessesPostgresIT starts two Main JVMs against a real PostgreSQL 17.6 server: an approval requested on A is decided on B, the approved delete runs once on B, the replay on A is refused, and a rate window of 3 is one window across both recorded local source behavior; see two-gateway evidence. Re-run from the exact candidate before a release claim
Knox extension API-level tests include the real AclsAuthz route-scope check; on 2026-10-05 a dated ZooKeeperFrontDoorsTest run installed that run’s knox-bundle jar (SHA-256 0726C4CE71D23856F8FE9A58BE2C80E7FC7E4AE740ED70E8ECC8CBFB2617B780) into stock Knox 3.0.0 ext/ and ran the full scenario against real ZooKeeper 3.9.3 (claims matrix H10). Knox 3.1 (Jakarta) needs a separate build, see ADR-002  

Every current configuration follows the same enforcement chain. Each profile names the version that was exercised. Later runs are on the roadmap.

Later work

Product-stack runs, typed sources, and the release-candidate repeat are on the roadmap.

Explicit limits

Reproduce

# From the repository root
C:\tools\apache-maven-3.9.11\bin\mvn.cmd -B clean install

# Adds real-engine modules where the profile can start an engine. Iceberg and Kafka require their
# loopback environment variables and servers before their tests execute.
C:\tools\apache-maven-3.9.11\bin\mvn.cmd -B "-Preal-engines" clean install

For each external target, use the source-specific setup and preservation instructions in the relevant tutorial and add the result to evidence/CLAIMS-MATRIX.md.